Privilege drift — the accumulation of permissions beyond an agent's original purpose — is a serious security risk in multi-agent systems (MAS). A FinOps Governance Platform example illustrates how five AI agents, each receiving incremental permission expansions over three months, collectively accumulate 20 API read scopes, 5 write surfaces, and access to sensitive financial, employee, and regulatory data. This expanded attack surface enables threats like prompt injection, over-privileged service accounts, lateral movement, and cascading authorization failures. Four practical mitigations are outlined: maintaining a platform-wide permission inventory, enforcing the principle of least privilege at runtime, implementing just-in-time (JIT) access with short-lived tokens, and continuously monitoring agent behavior for anomalies.
